Parallelization: run LLM calls side by side, then combine

Some work doesn't need to happen in sequence — it needs to happen at the same time. Parallelization fans a task out into independent LLM calls that run concurrently, then aggregates their outputs. It comes in two flavours, and both are useful for different reasons.

Two flavours

• Sectioning — split a task into independent subtasks (summarise each chapter, review each file) and run them at once, then stitch the results together.
• Voting — run the same prompt several times and aggregate by majority or threshold, trading tokens for reliability on high-stakes calls.

Sectioning wins on latency: three subtasks that would take 6s in sequence finish in ~2s in parallel. Voting wins on confidence: asking the model the same thing five times and taking the consensus catches the occasional bad roll.

const parts = await Promise.all([
  llm(prompt, sectionA),
  llm(prompt, sectionB),
  llm(prompt, sectionC),
]);

return aggregate(parts);   // merge, or take the majority vote

When to use it

Use sectioning when subtasks are genuinely independent and latency matters. Use voting when a single call is too unreliable for the stakes — content moderation, code review, anything where a false negative is expensive. The cost is more tokens; the payoff is speed or confidence.

If the subtasks don't depend on each other, waiting for them one at a time is just latency you chose to keep.